Abstract
We introduce a Gradient-weighted Class Activation Mapping (Grad-CAM) methodology to assess the performance of five distinct models for binary classification (normal/abnormal) of synchronized heart sounds and electrocardiograms. The applied models comprise a one-dimensional convolutional neural network (1D-CNN) using solely ECG signals, a two-dimensional convolutional neural network (2D-CNN) applied separately to PCG and ECG signals, and two multimodal models that employ both signals. In the multimodal models, we implement two fusion approaches: an early fusion and a late fusion. The results indicate a performance improvement in using an early fusion model for the joint classification of both signals, as opposed to using a PCG 2D-CNN or ECG 1D-CNN alone (e.g., ROC-AUC score of 0.81 vs. 0.79 and 0.79, respectively). Although the ECG 2D-CNN demonstrates a higher ROC-AUC score (0.82) compared to the early fusion model, it exhibits a lower F1-score (0.85 vs. 0.86). Grad-CAM unveils that the models tend to yield higher gradients in the QRS complex and T/P-wave of the ECG signal, as well as between the two PCG fundamental sounds (S1 and S2), for discerning normalcy or abnormality, thus showcasing that the models focus on clinically relevant features of the recorded data.

Abstract
In this work, we assess the impact of self-supervised learning (SSL) approaches on the detection of gastritis atrophy (GA) and intestinal metaplasia (IM) conditions. GA and IM are precancerous gastric lesions. Detecting these lesions is crucial to intervene early and prevent their progression to cancer. A set of experiments is conducted over the Chengdu dataset, by considering different amounts of annotated data in the training phase. Our results reveal that, when all available data is used for training, SSL approaches achieve a classification accuracy on par with a supervised learning baseline, (81.52% vs 81.76%). Interestingly, we observe that in low-data regimes (here represented as retaining only 12.5% of annotated data for training), the SSL model guarantees an accuracy gain with respect to the supervised learning baseline of approximately 1.5% (73.00% vs 71.52%). This observation hints at the potential of SSL models in leveraging unlabeled data, thus showcasing more robust performance improvements and generalization. Experimental results also show that SSL performance is significantly dependent on the specific data augmentation techniques and parameters adopted for contrastive learning, thus advocating for further investigations into the definition of optimal data augmentation frameworks specifically tailored for gastric lesion detection applications.

Abstract
Early diagnosis of cardiovascular diseases is essential for an effective treatment, potentially preventing severe health complications and improving clinical outcomes. Electrocardiogram (ECG) and phonocardiogram (PCG) are cost-effective, noninvasive diagnostic tools providing crucial and complementary information about the heart's electrical and mechanical activities. This paper presents a novel approach to the assessment of cardiovascular health through the multimodal analysis of simultaneously recorded ECG and PCG signals. Combining multimodal analysis and transfer learning on publicly available data, the most successful multimodal approach achieved an accuracy of 82.79%, a ROC AUC score of 91.26%, and a recall of 93.10% demonstrating the potential of these techniques. This study provides a foundation for future research aimed at enhancing the performance of multimodal cardiac abnormality detection systems. © 2024 IEEE.

Abstract
The detection of Pulmonary Hypertension (PH) from the computer analysis of digitized heart sounds is a low-cost and non-invasive solution for early PH detection and screening. We present an extensive cross-domain evaluation methodology with varying animals (humans and porcine animals) and varying auscultation technologies (phonocardiography and seisomocardiography) evaluated across four methods. We introduce PH-ELM, a resource-efficient PH detection model based on the extreme learning machine that is smaller (300x fewer parameters), energy efficient (532x fewer watts of power), faster (36x faster to train, 44x faster at inference), and more accurate on out-of-distribution testing (improves median accuracy by 0.09 area under the ROC curve (auROC)) in comparison to a previously best performing deep network. We make four observations from our analysis: (a) digital auscultation is a promising technology for the detection of pulmonary hypertension; (b) seismocardiography (SCG) signals and phonocardiography (PCG) signals are interchangeable to train PH detectors; (c) porcine heart sounds in the training data can be used to evaluate PH from human heart sounds (the PH-ELM model preserves 88 to 95% of the best in-distribution baseline performance); (d) predictive performance of PH detection can be mostly preserved with as few as 10 heartbeats and capturing up to approximately 200 heartbeats per subject can improve performance.

Abstract